Method and device for identifying an event in a field of view of an array of photovoltaic devices

Abstract

A computer implemented method for identifying an event in a field of view of an array of photovoltaic devices comprises: receiving, by processing circuitry, electrical parameters from one or more photovoltaic devices in the array, wherein the electrical parameters are dependent on electromagnetic wave propagation in the field of view of the array of photovoltaic devices; wherein the processing circuitry comprises a trained machine learning model configured to identify an event taking place in the field of view of the array based on the electrical parameters; and using the trained machine learning model to identify an event.

Description

[0001] Method and Device for Identifying an Event in a Field of View of an Array of Photovoltaic Devices

[0002] Field

[0003] The disclosure relates to a method and a device for identifying an event in a field of view of an array of photovoltaic devices. Examples described herein relate generally to a method and device for processing data, for example for training and using a machine learning model to provide event identification.

[0004] Background

[0005] Optical neural networks are an appealing physical realization of digital neural networks and perform part of or the entirety of a computation using light. These computations can often be done in a lower-cost I cost-free manner, which can be leveraged to create a faster, more efficient and more sustainable implementation of their digital counterparts.

[0006] Typical implementations of optical neural networks (ONN) comprise an optical source such as a laser, a method to shape the illumination such as a spatial light modulator, free space or an engineered medium through which light propagates and a detector such as a CCD camera / sensor. A network of single-pixel light detectors or cameras may also be used as the detector. All ONN implementations require a degree of nonlinearity in the path between the source and the detector and benefit from a large number of pixels in the detector.

[0007] The creation of a distributed network of bespoke optical detectors is not scalable and it is not conceivable to drive a capillary distribution of the CCD sensors. Additionally, the use of high-resolution cameras would create considerable concerns for the privacy and safety of individuals and institutions.

[0008] To achieve non-linearity, light detectors may be driven beyond their linear regime, to obtain a nonlinear response in the network, but this is not ideal as it would require dedicating expensive equipment to the ONN system. “Image sensing with multilayer nonlinear optical neural networks, Tianyu Wang et. al.” describes a multilayer optical neural network (ONN) encoder with non-linearity for image sensing which uses a commercial image-intensifier as an optical-to-optical nonlinear activation function. “Single-shot polarimetry of vector beams by supervised learning, Davide Pierangeli et. al.” describes polarimetry of vector beams in a single shot without the use of any polarization optics. “Dynamic recognition and mirage using neurometamaterials, C. Quian et. al.” describes the use of neuro-metamaterials to realize a dynamic object-recognition system.

[0009] Summary

[0010] One or more aspects of the present disclosure relate to a computer-implemented method for identifying an event in a field of view of an array of photovoltaic devices. The method comprises, as a first step, receiving, by processing circuitry, electrical parameters from one or more (e.g. multiple, a plurality of or each) photovoltaic device in the array, wherein the electrical parameters are dependent on electromagnetic wave propagation in the field of view of the array of photovoltaic devices. The processing circuitry comprises a trained machine learning (ML) model configured to identify an event taking place in the field of view of the array based on the electrical parameters. In a second step, the trained machine learning model is used to identify an event.

[0011] The machine learning model may comprise an artificial neural network (ANN). In some examples, the machine learning model may comprise a regression model, support vector machine or k-nearest-neighbour (KNN) model. In other examples, the machine learning model may comprise a K-means or spectral clustering model.

[0012] One aspect of the disclosure relates to a machine learning model which receives signals from photovoltaic devices. In some examples, the photovoltaic devices may be solar cells. Each photovoltaic device may constitute a pixel of a massively distributed detection system, trained to identify and / or track events in real-time. Thus, the system is akin to an optical neural network. The processing is performed initially by the propagation and detection of light and subsequently by processing circuity comprising a machine learning model. The processing circuitry may be in the form of hardware and / or software. The system has the ability to access in real-time the photovoltage and / or photocurrent generated in the photovoltaic devices. The system may make use of existing installed infrastructure, both at the industrial and private user levels. The system is independent of the type of photovoltaic device used.

[0013] Solar cells are ubiquitous, and they already cover substantial areas, with capillary distributions. Solar cells do not have the ability to create high-resolution images, thus drastically reducing issues related to security and privacy. Solar cell installations are ideal as large-area detectors, e.g. for satellite communications.

[0014] Solar cells have an intrinsic nonlinear behaviour, directly accessible by obtaining the photovoltage generated by solar illumination. The photovoltage, photocurrent or parameters derived therefrom are used as the input signal for the machine learning model. These parameters are readily accessible at the point of generation. Solar cellbased ANNs offer a uniquely sustainable implementation of neural networks, as they may produce their own power using solar energy and may not need external power sources (although, in some cases, additional power sources may be employed).

[0015] Solar cell technology is well-suited for creating high quality neural networks wherein the distribution is capillary. No new large-scale infrastructures may be required. The system may extend neural network capability to large scale events and satellite and Li-Fi optical communication systems. There is an opportunity to develop driving / monitoring optoelectronic interfaces to be applied to existing solar cells farms and private end user owned solar cells.

[0016] The network architecture can be adapted to detect events such as meteorological events, bird migration patterns, or for aerial surveillance, or optical communication such as LiFi and satellite communication. The optical source in the above applications may be light from the sun or from an artificial source, the medium is the natural environment and a suitable number of detectors in the form of photovoltaic devices can be connected to complete the network. For the specific case of long-distance optical communication, e.g. for Li-fi and Satellite / ground communications, it is beneficial to use large area detectors, which are difficult to source off-the-shelf.

[0017] One aspect of the disclosure comprises processing circuitry comprising a machine learning model which receives the electrical parameters from the photovoltaic devices and uses them as inputs to the machine learning model. In another aspect an interfacing device is presented that may interface with existing photovoltaic device installations to obtain the required electrical parameters in real-time, such as generated photovoltage, and transmit these parameters to the machine learning model. The interfacing device may measure the electrical parameters associated with the photovoltaic devices. The obtaining or the measuring of the electrical parameters may be achieved by using interface circuitry of the interfacing device. The processing circuitry, which comprises the model, and the interfacing circuitry may be provided together in a processing and interfacing device or apparatus. The processing and interfacing device may further include an output device to convey a notification of an identified event, received from the processing circuitry. In some cases, the interfacing device may include only the interface circuitry and may be configured to provide the electrical parameters to processing circuitry in a separate processing device.

[0018] The system may implement a new type of optical edge computing based on machine learning models, which are typically used for image classification, regression problems, and the like, where the processing of the information happens close to the point at which the information is generated.

[0019] Examples of the first aspect of the disclosure therefore relate to a method of identifying an event using a machine learning model that is trained to use electrical parameters from arrays of photovoltaic devices as inputs. Advantageously, the photovoltaic devices may be constituted by solar cells and existing infrastructure may be utilised to generate the electrical parameters required by the processing circuitry.

[0020] The electrical parameters obtained from the photovoltaic devices are representative of electromagnetic signals received by the photovoltaic devices, the electromagnetic signals being received from within the field of view of one or more (e.g. multiple, a plurality of or each) photovoltaic device. The model is trained to identify events in the fields of view of the photovoltaic devices and receives the electrical parameters or derivatives thereof as input while providing the identity of at least one event as an output. The output may be displayed on an output device or stored in a memory device, which may or may not be provided within the processing circuitry.

[0021] In another example, the method includes using interface circuitry to obtain electrical parameters from the array of photovoltaic devices and to provide the electrical parameters to the processing circuitry. The interface circuitry may be used to obtain the electrical parameters from the photovoltaic devices and to provide the electrical parameters to the processing circuitry. The interface circuitry may measure the electrical parameters of the photovoltaic devices. The interface circuitry may comprise the circuitry necessary for measuring the electrical parameters of the photovoltaic devices.

[0022] The electrical parameters may comprise one or more of measured voltage, current and output power obtained from the photovoltaic device. The electrical parameters received / measured may comprise voltage, current or derivatives thereof from any electrical circuitry associated with the photovoltaic devices. The parameters may comprise photovoltage, photocurrent and derivatives thereof. The electrical parameters may be measured over a period of time as events elapse in the field of view of the photovoltaic devices. For example, the electrical parameters may be measured (e.g. sampled) continuously, near-continuously, periodically or according to a pre-defined schedule or series of intervals.

[0023] The field of view of the array may comprise one or more solid angles through which the array can detect electromagnetic wave propagation. The field of view of a photovoltaic device is the solid angle through which the photovoltaic device is sensitive to electromagnetic radiation. The field of view of a solar cell, for example, is the solid angle in space from which sunlight can be detected when incident on the photovoltaic cell. Fields of view of individual photovoltaic devices may be generally conical having a vertex disposed on a detector element and an axis perpendicular to a detection surface of the detector element. However, the shape and orientation of the field of view may be dependent on the construction of the detector element and may vary for different kinds of photovoltaic devices. The fields of view of individual photovoltaic devices may overlap. This is especially true for the case of a planar array where the axes of the fields of view may align in parallel. An amount of overlap may then depend on the physical separation of the photovoltaic devices (for example, between rows and columns of an array). An array of photovoltaic devices disposed in proximity to each other may take advantage of a combined field of view comprising the individual fields of view of one or more (e.g. multiple, a plurality of or each) photovoltaic device and an amount of overlap (which may be relatively large, e.g. 50% or more). The field of view of the array of photovoltaic devices is the solid angle in space which may comprise the combined fields of view of some or all of the photovoltaic devices, especially when the individual fields of view overlap. In some examples, the array comprises a plurality of photovoltaic devices arranged in one or more directions. The array may comprise one photovoltaic device or a plurality of photovoltaic devices arranged in one direction (e.g. a line of photovoltaic devices may constitute an array). The array of photovoltaic devices may be arranged in a two- dimensional planar array with each photovoltaic device constituting an array element. The photovoltaic devices in the array may be electrically connected to each other. The fields of view of the photovoltaic devices in the array may overlap substantially (e.g. by 50% or more). In some cases, the fields of view of the photovoltaic devices may not overlap at all or may only overlap for a small solid angle (e.g. 10% or less) in comparison to the solid angle for an individual photovoltaic device.

[0024] In some cases, the array may be three-dimensional or may comprise a single photovoltaic device or array element. In some cases, the array elements may be spread out geographically. In such a case, each element of the array may comprise a single photovoltaic device or a plurality of photovoltaic devices within a geographical area with a combined field of view that comprises the individual fields of view of all the photovoltaic devices in the array. In other words, the array may comprise multiple photovoltaic devices provided in a single location (e.g. to form a single array) or individual photovoltaic devices provided in multiple locations (e.g. to form a distributed array) or groups of photovoltaic devices provided in multiple locations (e.g. to form an array of arrays); or any combination thereof.

[0025] The event to be identified in the field of view of the array may comprise one or more of a meteorological event, movement of birds, an aerial surveillance event, a lidar event, optical communication and optical satellite communication. The meteorological event may comprise the presence or the absence of cloud cover in the sky, or the proportion of cloud cover. The identification of the meteorological event may comprise the identification of a clear sky with low or no cloud cover. The meteorological event may comprise snow, rain, hail, sleet and other such weather-related events that are identifiable in the field of view of the array. The movement of birds may be the movement of one or more birds through the field of view of the array. Individual species of birds may be identifiable in some examples. Aerial surveillance events may comprise the movement or presence of aircraft, weather balloons, satellites, space debris and other physical items in the field of view of the array. In some examples, the processing circuitry stores at least an indication of the event in a memory and / or provides an identification of the event to the user in the form of a notification on an output device. The memory where the indication of the event is stored may be a physical memory device. The memory device may be included in the processing circuitry or may be provided separately. In some cases, the memory may be hosted on hardware that is not physically connected to the processing circuitry, such as on the cloud or hosted at a central server physically removed from the processing circuitry and the array. In such cases, the processing circuitry may transmit the indication of the event to the memory using an appropriate communication method such as wired or wireless communication. The output device may be a computer monitor or mobile device and the indication may be a visual indication. In some examples, the output device may be a speaker and the indication may be an aural indication. Other variations and combinations of output indications may be used in other cases.

[0026] In another aspect, an apparatus is presented for identifying an event in a field of view of an array of photovoltaic devices comprising processing circuitry configured to receive electrical parameters from the array, wherein the electrical parameters are dependent on electromagnetic wave propagation in the field of view of the array of photovoltaic devices, the processing circuitry comprising a trained machine learning model configured to identify an event taking place in the field of view of the array based on the electrical parameters. The apparatus may comprise a physical component such as hardware and a virtual component such as software. The apparatus may be comprised within a physical device or be distributed over several physical devices. In some cases, the apparatus may comprise a physical device collocated with the array as well as a virtual or software component physically removed from the physical device and the array. The device may comprise a communication component for the transfer of data between a physical device and a virtual or software component.

[0027] In some examples, the apparatus comprises interface circuitry configured to obtain the electrical parameters from one or more photovoltaic devices in the array of photovoltaic devices and to provide the electrical parameters to the processing circuitry. In some examples the interface circuitry is further configured to measure the electrical parameters from the array of photovoltaic devices. The interface circuitry may comprise the circuitry required to obtain or measure the electrical parameters from the array of photovoltaic devices. The interface circuitry may be able to interface with one or more (e.g. multiple, a plurality of or each) photovoltaic device in an array of photovoltaic devices to obtain / measure the electrical parameters.

[0028] In another aspect, a system is presented, the system comprising: one or more arrays of photovoltaic devices; and processing circuitry configured to receive electrical parameters from photovoltaic devices in the one or more arrays, wherein the electrical parameters are dependent on electromagnetic wave propagation in the fields of view of the arrays of photovoltaic devices; wherein the processing circuitry comprises a trained machine learning model configured to identify an event taking place in the fields of view of the one of more arrays based on the electrical parameters.

[0029] In some examples, the system further comprises interface circuitry configured to obtain and / or measure electrical parameters from one or more photovoltaic devices in the one or more arrays of photovoltaic devices and to provide the electrical parameters to the processing circuitry. The interface circuitry may comprise the circuitry / hardware / software required to obtain or measure the electrical parameters from the one or more arrays of photovoltaic devices. The interface circuitry may be able to interface with one or more (e.g. multiple, a plurality of or each) photovoltaic device in an array of photovoltaic devices as well as a plurality of such arrays to obtain / measure the electrical parameters.

[0030] In some examples, the interface circuitry is further configured to determine one or more of a time associated with the received electrical parameters, the location of the one or more arrays of photovoltaic devices associated with the received electrical parameters and the orientation of the one or more arrays of photovoltaic devices associated with the received electrical parameters. The interface circuitry may have the ability to determine one or more time values associated with the received electrical parameters. Each reading of electrical parameters may be accompanied by a time stamp denoting when the reading was taken. The interface circuitry may further comprise geolocation circuitry such as the Global Positioning System (GPS). The geolocation circuitry may be able to measure the location of the photovoltaic devices associated with the interface circuitry. In some examples, the time stamps associated with electrical parameter readings may be accompanied by the location co-ordinates of the photovoltaic device or devices. The interface circuitry may further comprise orientation circuitry such as an accelerometer to measure the orientation of the photovoltaic device or array of photovoltaic devices. The orientation measured may be indicative of the axis of the field of view of the device or array. In some examples, the measured orientation of the device or array may accompany readings of electrical parameters. In some examples, each reading of electrical parameters may be accompanied by a time, location and orientation so measured. In some cases, the location and / or orientation of the device and / or array may be pre-determined and stored in a memory accessible by the processing circuitry. An identifier may be associated with each device and / or array and the electrical parameters obtained from each device and / or array may be associated with a corresponding identifier.

[0031] In some examples, the system comprises a plurality of photovoltaic device arrays, wherein a plurality of the processing circuitries and a plurality of interface circuitries are provided and each one of the plurality of interface circuitries is configured to provide the electrical parameters to a respective one of the plurality of the processing circuitries. In some examples, the system comprises a plurality of photovoltaic device arrays, wherein one or more of the processing circuitries and a plurality of the interface circuitries are provided and the plurality of interface circuitries are configured to provide the electrical parameters to the one or more processing circuitries. In further examples, the system may comprise a plurality of photovoltaic device arrays, wherein a plurality of the interface circuitries is provided and the plurality of interface circuitries is configured to provide the electrical parameters to a number of processing circuitries wherein the number of processing circuitries is smaller than a number of interface circuitries. In this way, the system may function in multiple different configurations. In the case where an array is associated with an interface circuitry and a processing circuitry, edge-computing may be implemented. In the case where a plurality of arrays is associated with as many interface circuitries but only one processing circuitry, a central-server may be implemented to host the processing circuitry. In such a case, the processing circuitry, which comprises the machine learning model may be hosted remotely from the location of the arrays and interface circuitries.

[0032] In some examples, the system may comprise processing circuitry configured to track events taking place in a combined field of view of a plurality of photovoltaic devices in the one or more arrays, spatially and / or temporally. The plurality of photovoltaic devices may comprise a plurality of arrays of photovoltaic devices. The plurality of photovoltaic devices may comprise a plurality of individual photovoltaic devices. In some cases, the plurality of photovoltaic devices may comprise a combination of arrays of photovoltaic devices and individual photovoltaic devices. In some examples, the photovoltaic devices comprise one or more solar cells such as Perovskite solar cells, single junction solar cells and organic solar cells. In some examples, at least one of the interface circuitries and the processing circuitries is powered by at least the photovoltaic devices. Other sources of power may also be used to power at least one of the interface circuitries and at least one of the processing circuitries. In some examples, at least one of the interface circuitries are powered completely by the photovoltaic devices. In some examples, at least one of the processing circuitries are powered completely by the photovoltaic devices.

[0033] In some examples, the system further comprises an output device configured to convey a notification of an identified event, received from the processing circuitry and / or received from the memory. The output device may be a computer monitor or mobile device and the indication may be a visual indication. In some examples, the output device may be a speaker and the indication may be an aural indication. Other variations and combinations of output indications may be used in other cases. In some examples, an indication of the event may be stored in a memory device.

[0034] In an aspect of the disclosure wherein the machine learning model is a supervised learning model, a processing apparatus configured to train a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices is presented.

[0035] The processing apparatus may comprise training circuitry configured to train the supervised machine learning model to perform event identification based on training data, wherein the training data comprises or is related to electrical parameters associated with events and wherein the electrical parameters are obtained from the array of photovoltaic devices. For example, the electrical parameters may be obtained from one or more (e.g. multiple, a plurality of or each) photovoltaic devices in one or more arrays. The processing apparatus may be provided within the processing circuitry. In some cases, the processing apparatus may be a separate device or software. The processing apparatus may comprise its own processing resources such as memory, a processor and peripherals. In some cases, the processing apparatus may share processing resources with the processing circuitry.

[0036] The machine learning model may comprise an artificial neural network (ANN). The ANN may comprise a convolutional neural network (CNN). The ANN may comprise a recurrent neural network (RNN). An RNN is considered to be particularly suitable for analysing temporal changes. The ANN may comprise a transformer, such as a vision transformer. The ANN may comprise a multilayer perceptron (MLP). A MLP is considered to be particularly suitable for a low number of optical inputs. The ANN may comprise a graph neural network (GNN). A GNN is considered to be particularly suitable for a large number of inputs. The ANN may perform one or more of image classification, image captioning, image segmentation, anomaly detection or action recognition on the basis of the input electrical parameters and / or derivatives thereof.

[0037] In an aspect of the disclosure, wherein the machine learning model is a supervised learning model, a method of training a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices is presented.

[0038] The method may comprise training the machine learning model to perform event identification based on training data, wherein the training data comprises or is related to electrical parameters associated with events and wherein the electrical parameters are obtained from the array of photovoltaic devices.

[0039] The method may comprise obtaining the electrical parameters associated with the fields of view of the one or more arrays of photovoltaic devices for use in training the model. The fields of view may be chosen by an operator to model the fields of view of interest based on the application envisioned. For example, when the application envisaged is meteorological, the fields of view may comprise clear skies, cloudy skies and other weather conditions such as the presence of rain, sleet or snow. In some examples, where the application envisaged is that of bird migration / movement, the fields of view used for the training data may comprise the movement of one or more birds in flight. When the application envisaged is that of aerial surveillance, the fields of view may comprise the movement of aircraft and other objects in flight. The fields of view may further be chosen to cover different times of the day or night to include variations in ambient light.

[0040] The electrical parameters obtained from one or more arrays of photovoltaic devices may be used to train the model. One or more of supervised learning, unsupervised learning and self-supervised learning may be used to train the model. Other methods of learning that the person skilled in the art would envisage may also be used. The electrical parameters may be processed by the processing apparatus to obtain data sets representative of electromagnetic wave propagation in the field of view of the one or more arrays of photovoltaic devices. The data sets may be in the form of arrays of electrical parameters corresponding to the physical arrangement of photovoltaic devices in the arrays where one or more (e.g. multiple, a plurality of or each) photovoltaic device forms a pixel of the array. The data sets in these cases would be similar to optical images received by the one or more arrays. These data sets may be used to train the model to identify events taking place in one or more fields of view of the one or more arrays of photovoltaic devices. In some examples, a human labeller with expert knowledge may label the electrical parameters obtained from the photovoltaic devices and corresponding to the presence of one or more of weather phenomena, birds and aerial objects, to provide the ground truth for training the ANN / ML model, The training of the model may depend on the location and / or orientation of the associated array.

[0041] In some examples, the model may be trained to interpolate movement of an event (e.g. a bird, cloud or aircraft) in the space and time between the event being captured in the field of view of one photovoltaic device or array and another photovoltaic device or array.

[0042] These and other aspects will be apparent from the examples described in the following. The scope of the present disclosure is not intended to be limited by this summary nor to implementations that necessarily solve any or all of the disadvantages noted.

[0043] Any features described in relation to one aspect of the disclosure may be applied to any one or more other aspect of the disclosure.

[0044] Brief description of the drawings

[0045] Examples are now described, by way of non-limiting example, and are illustrated in the following figures, in which:

[0046] Figure 1 is a flowchart illustrating a method in accordance with the disclosure;

[0047] Figure 2 is a schematic of a system in accordance with the disclosure;

[0048] Figure 3 is a schematic of a system in accordance with the disclosure;

[0049] Figures 4a and 4b are illustrations of different fields of view of photovoltaic devices in accordance with the disclosure;

[0050] Figure 5 is a schematic of a system in accordance with the disclosure; Figure 6 is a schematic of a system in accordance with the disclosure;

[0051] Figure 7 is a graph showing the response of a Perovskite solar cell sample to input light power in accordance with the disclosure;

[0052] Figure 8 is a schematic of an apparatus used to test the accuracy of a method in accordance with the disclosure; and

[0053] Figure 9 is a violin graph comparing the accuracy of two different set-ups during testing.

[0054] Detailed description

[0055] Figure 1 is a flowchart illustrating a computer implemented method 100 for identifying an event in a field of view of an array of photovoltaic devices. In step 2 of the method 100, electrical parameters are obtained from one or more (e.g. multiple, a plurality of or each) photovoltaic device in an array of photovoltaic devices. The electrical parameters may include voltage, current and / or derivative parameters thereof. The electrical parameters may be accompanied by one or more of the time, the location of the photovoltaic device and / or array and the orientation of the photovoltaic device and / or array when the reading was taken. The electrical parameters may relate to individual photovoltaic devices or may relate to a collection of photovoltaic devices.

[0056] The electrical parameters are dependent on electromagnetic wave propagation in the field of view of the array. The field of view is the solid angle through which the photovoltaic devices of the array are sensitive to electromagnetic radiation. Electromagnetic radiation incident on the array that falls within the field of view is detected by the array. Variations in the medium contained within the field of view may result in variations in the electromagnetic energy received by the array. Such variations further result in variations in the electrical parameters. The electrical parameters are hence representative of the state of electromagnetic wave propagation in the field of view of the array. Events such as meteorological events can cause such variations. The photovoltaic devices may comprise solar cells. For such cases, the electrical parameters obtained while the sky is bright and clear of clouds may be averaged to serve as a baseline to compare with electrical parameters received when an event is taking place (e.g. when a cloud or other obstruction is in the sky). In some cases, other baselines may be used, for example, when the sky is completely dark. The use of baselines allows the system to identify events in the case of an absence of variations in the medium contained within the field of view. In one example, the method may detect the presence of clouds over a period of time without any significant variation in cloud cover. Similarly, a persistently clear sky could be identified despite the lack of electromagnetic variations that contrast a clear sky from a cloudy sky. As such, the identification of an event may not be limited to discrete events, which involve a change in the electrical parameters. The identification of an event may comprise identifying a persistent event where the electrical parameters are substantially unchanged. The presence of a persistent event may indicate a lack of a discrete event. In some cases, the system may be configured to identify a lack of an event of interest by, for example, identifying a different event to the event of interest. For example, a lack of sunshine may be determined by identification of a cloudy day.

[0057] Interface circuitry may be used to obtain the electrical parameters from the photovoltaic devices. The electrical parameters associated with photovoltaic devices may be measured by the interface circuitry. The interface circuitry may comprise clock circuitry, geolocation circuitry and orientation circuitry. The electrical parameters obtained may be accompanied by data obtained from one or more of the clock circuitry, the geolocation circuitry and the orientation circuitry. For example, the clock circuitry may determine the time at which the electrical parameters were measured; the geolocation circuitry may determine the location of the photovoltaic devices when the electrical parameters were measured; and the orientation circuitry may determine the orientation of the photovoltaic devices when the electrical parameters were measured. In some cases, the location and / or orientation data may be stored in a memory accessible by the processing circuitry. An identifier may be associated with each photovoltaic device and / or array and the electrical parameters obtained from each photovoltaic device and / or array may be associated with a corresponding identifier. Thus, the location and / or orientation of each photovoltaic device and / or array may be determined using the identifier to query the memory in order to relay the location and / or orientation data along with the electrical parameters and time to the processing circuitry.

[0058] Events to be identified may include meteorological events such as cloud cover, rain, fog, mist or other events that change the characteristics of the medium (air in this case) with respect to electromagnetic radiation and cause a change in electrical parameters obtained from the array. In some cases, events may comprise the movement of birds in the field of view of the array. This may not be limited to birds but may be extended to any such change that is registrable with the detector, such as the movement of airplanes, satellites, and a variety of other aerial surveillance subjects. An event may comprise the reception of electromagnetic radiation of a particular type or configuration. Any photovoltaic device used in the array will have a frequency response that defines its ability to absorb electromagnetic radiation of a range of frequencies. Since the frequency response of photovoltaic devices comprises at least the optical range of frequencies, the photovoltaic devices may be used to receive signals associated with LiFi, lidar, optical communication and optical satellite communication.

[0059] In step 4 of the method 100, the electrical parameters are provided to the processing circuitry. The parameters may be relayed to the processing circuitry via the interface circuitry. In some cases, the electrical parameters may be provided directly to the processing circuitry from the array of photovoltaic devices. In such cases, the processing circuitry may be collocated with the array and may include the time circuitry, the geolocation circuitry and the orientation circuitry mentioned above. The processing circuitry comprises an ANN trained to identify events in the field of view of the array of photovoltaic devices, based on the electrical parameters.

[0060] In step 6 of method 100, the ANN is used to identify one or more events in the field of view of the array.

[0061] The ‘identification step’ (i.e. inferencing process) comprises performing feedforward propagation for the inputs to the ANN, and obtaining the model outputs.

[0062] To enable identification, the ‘training step’ is crucial. AN Ns may generally be trained in three settings: supervised (with ground truth provided to the model), unsupervised (using models trained without the access to the ground truth) and reinforcement learning (where the problem is framed as a series of decision making steps, and trained by assigning multiple rewards at the final or some intermediate states of the decision making steps). Any suitable training method may be employed.

[0063] In examples where another machine learning algorithm is used instead of an ANN, the input of the photovoltaic devices and a multilayer (for example, four layer) perceptron (MLP) may be used to identify events in the field of view of the array. A cross-entropy loss can be computed at the output layer (last layer of the MLP), and the gradient of which is back-propagated through each of the nodes within the MLP to update the weights associated with each node accordingly. In some examples, batch normalisation, gradient clipping and one of many regularisations may be used as would be understandable to the person skilled in the art.

[0064] The output for image classification is often a number or an array of numbers, each corresponding to at least one, or a combination of possible outcomes. The final result is taken by computing the argmax function in the case of an array of numbers, which corresponds to a pre-defined event or to predefined events. In the case where the model output is only one number, it may be used for the identification of binary contradicting events such as bright or dark etc.

[0065] Integration I interfacing with other models is also possible. For example, the collected data can be stored as latents, that are not directly interpretable but still possess sufficient information to be used for downstream tasks I pipelines, such as the automated decision making of changing the orientation of certain type of light- or event-sensitive instruments or in the decision-making of a self-driving car. These are sometimes used because it allows for more efficient storage of data, and allows the ML model to be trained end-to- end as a single model instead of two or more separate systems wherein each may introduce its own errors or inaccuracies.

[0066] It is possible to use one or more trained auto-encoders in some examples. The autoencoders may be used to map the inputs to a latent space, which can be augmented by text, audio or other encoded inputs of other modalities to generate an output accordingly.

[0067] Or one may consider any sort of integration as a two-part system, where a first part identifies the event and passes the result to a downstream model.

[0068] In step 8 of method 100, an output device may be used to convey a notification of an identified event, received from the processing circuitry.

[0069] Figure 2 is a schematic of a system 200 used to identify an event in a field of view of an array of photovoltaic devices. An array 12 of photovoltaic devices is used to detect electromagnetic radiation incident on it through its field of view. The array 12 comprises a plurality of photovoltaic devices 20. The array 12 is a planar array with the photovoltaic devices 20 disposed substantially in one two-dimensional plane. The array 12 is provided in an outdoor environment with a clear view of the sky. In some examples, the array 12 may comprise two or more individual photovoltaic devices provided in different locations. In this case, the photovoltaic devices 20 are solar cells. The photovoltaic devices may be Perovskite solar cells, single junction solar cells, organic solar cells or a combination of two of more of these types of cells. In some cases, the array 12 may be provided in an indoor environment.

[0070] Electrical parameters associated with one or more (e.g. multiple, a plurality of or each) photovoltaic device 20 are provided to interface circuitry 14. In some cases, the interface circuitry 14 includes measurement circuitry that measures the electrical parameters from the photovoltaic devices 20 in the array 20. The interface circuitry 14 may comprise one or more of clock circuitry, geolocation circuitry and orientation circuitry. At least part of the interface circuitry 14 is collocated with the array 12. The interface circuitry 14 is electrically connected to the array 12.

[0071] The interface circuitry 14 provides the electrical parameters to the processing circuitry 16. The processing circuitry 16 may receive data from one or more data stores (not shown) instead of or in addition to the interface circuitry 14. For example, the processing circuitry 16 may receive data from one or more remote data stores (not shown) which may form part of a Picture Archiving and Communication System (PACS) or other information system.

[0072] The processing circuitry 16 comprises a computing apparatus, which may comprise a personal computer (PC) or workstation. The computing apparatus may be connected to a display screen or other display device, and an input device or devices, such as a computer keyboard and mouse. In some cases, the computing apparatus may be cloudbased.

[0073] The processing circuitry 16 provides a processing resource for automatically or semi- automatically processing data. The processing circuitry 16 may comprise a processing apparatus configured to train a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices, the processing apparatus configured to train one or more models and I / O circuitry configured to obtain user or other inputs and output results. The I / O circuitry may comprise an output device 18 configured to convey a notification of an identified event, received from the processing circuitry and / or received from the memory 17. The output device may be a computer monitor and the indication may be a visual indication. In some examples, the output device may be a speaker and the indication may be an aural indication. Other variations and combinations of output indications may be used in some cases. The output device may provide an indication about the type of event, severity of the event (such as for meteorological events), duration of the event, location of the event and other information about the event identified. The I / O circuitry may comprise an input device (not shown) through which a user may specify a type of event of interest and the processing circuitry 16 may output information relating to an identification of the type of event of interest to the user via the output device 18.

[0074] In the present case, the processing circuitry 16 is implemented in a computing apparatus by means of a computer program having computer-readable instructions that are executable to perform the method described. However, in some cases, the various circuitries, including the processing circuitry 16 and the interface circuitry 14 may be implemented as one or more ASICs (application specific integrated circuits) or FPGAs (field programmable gate arrays).

[0075] The computing apparatus also includes a hard drive and other components of a PC including RAM, ROM, a data bus, an operating system including various device drivers, and hardware devices including a graphics card.

[0076] The connection between the interface circuitry 14 and the processing circuitry 16 may be electrical. In some cases, the connection may rely on a wireless transmission of data from the interface circuitry 14 to the processing circuitry 16. The processing circuitry 16 may be collocated with the interface circuitry 14 and the array 12 or remote therefrom. The processing circuitry 16 may comprise electronic hardware. In some cases, the processing circuitry 16, or at least a portion of it, may be cloud-based. In examples wherein at least the array 12, interface circuitry 14 and processing circuitry 16 are collocated, the system 200 may be considered to constitute an example of edge computing whereby the processing is performed at the site of the array 12.

[0077] The processing circuitry 16 comprises a ML model trained to identify one or more events in a field of view of the array 12. The processing circuitry 16 provides the electrical parameters or derivatives thereof to the ML model as inputs thereto. The ML model analyses the inputs to identify one or more events and, once identified, conveys a notification of an identified event as an output, which may be transmitted from the processing circuitry 16 to an output device 18. In some cases, the notification of an identified event is stored in memory 17 and the output device 18 is not required as part of the system 200. Figure 3 is a schematic of a system 300 used to identify an event in a field of view of an array 22 of photovoltaic devices 20. Features described in relation to of Figure 2 will not be detailed in the description of the system 300. The description will only emphasise the differences between the system 300 and the system 200. The system 300 differs from the system 200 in that it does not include the interface circuitry 14. The array 22 is electrically connected to the processing circuitry 26 in a direct manner. The processing circuitry 26 may comprise one or more of measuring circuitry, clock circuitry, geolocation circuitry and orientation circuitry. The processing circuitry 26 obtains the electrical parameters from the array 22 and provides them to the ML model trained to identify one or more events in the field of view of the array 22. The processing circuitry 26 transmits a notification of an identified event, received as output from the ML model, to the output device 28. The processing circuitry in this case is collocated with the array 22. As such, the system 300 may be considered to constitute an example of edge computing whereby the processing is performed at the site of the array 22.

[0078] Figures 4a and 4b are illustrations of different fields of view of photovoltaic devices in accordance with the disclosure. A field of view of the array comprises one or more solid angles through which the array can detect electromagnetic wave propagation. The field of view of a photovoltaic device is the solid angle through which the photovoltaic device is sensitive to electromagnetic radiation. Figures 4a and 4b show two-dimensional renderings of conical three-dimensional fields of view 21 of the photovoltaic devices 20. Only two photovoltaic devices are shown in each of Figures 4a and 4b for the purposes of illustration. The concepts discussed with regard to Figures 4a and 4b can be extended to any number of photovoltaic devices 20 and their associated fields of view 21 .

[0079] Figure 4a shows two photovoltaic devices 20 and their associated fields of view 21. The two-dimensional renderings of the conical fields of view 21 are illustrated as inverted triangles having vertices disposed on each photovoltaic device 20 and an axis perpendicular to a detection surface of a detection element of the photovoltaic device 20. In some cases, the shape and orientation of the fields of view 21 may be dependent on the construction of the detector element and may vary for different kinds of photovoltaic devices 20. The fields of view 21 shown in Figure 4a are shown not to overlap. Figure 4b shows two photovoltaic devices 20 and their associated fields of view 21 . The arrangement shown in Figure 4b differs from that of Figure 4a in that the fields of view 21 of the photovoltaic devices 20 in Figure 4b overlap, leading to the creation of a field of view overlap region 23. The overlap region 23 comprises one or more solid angles through which both photovoltaic devices 20 can detect electromagnetic wave propagation or are sensitive to electromagnetic radiation. The fields of view 21 in Figure 4b could overlap for a variety of reasons, such as the photovoltaic devices 20 being placed closer together than in Figure 4a, the vertex angles of the fields of view being larger or the photovoltaic devices 20 being sensitive to electromagnetic radiation at a greater distance than the photovoltaic devices 20 of Figure 4a, among others.

[0080] Figure 5 is schematic of a system 400 used to identify an event in a field of view of an arrays 30, 31 , 36 of photovoltaic devices 20 and is independent of the previous systems described above. Features described in relation to the previous systems will not be described in detail in the description of the system 400. System 400 comprises a plurality of arrays 30, 31 , 36 of photovoltaic devices 20. While only three arrays are shown in Figure 5, the system 400 may comprise more than three arrays 30, 31 , 36. Each array

[0081] 30, 31 , 36 may comprise a plurality of photovoltaic devices 20 or only one photovoltaic device 20. The arrays 30, 31 , 36 may have overlapping fields of view as illustrated in Figure 4b or their fields of view may be separated (or adjacent) so that there is no overlap as shown in Figure 4a. The arrays 30, 31 , 36 may be located close to each other or be any distance away from each other.

[0082] Each array 30, 31 , 36 of system 400 is connected to a, respective interface circuitry 32,

[0083] 33, 37. The interface circuitries 32, 33, 37 are collocated with the respective array 30,

[0084] 31 , 36 and are electrically connected to the respective array 30, 31 , 36. The interface circuitries 32, 33, 37 obtain electrical parameters from the respective arrays 30, 31 , 36 and provide them to respective processing circuitries 34, 35. The processing circuitries

[0085] 34, 35 may be collocated with the respective array 30, 31 , 36 and the respective interface circuitries 32, 33, 37 or may be remote therefrom. The connection between the interface circuitries 32, 33 and the processing circuitries 34, 35 may be a wired connection. In some cases, the connection may be wireless. In this case, each array 30, 31 is associated with at least one interface circuitry 32, 33 and at least one processing circuitry 34, 35. Arrays 31 and 36 are associated with their respective interface circuitries 33, 37 but only one processing circuitry 35 in order to illustrate the different variations possible. Each processing circuitry 34, 35 may comprise a ML model trained to identify one or more events in the field of view of the respective array 30, 31 , 36. The ML models hosted in each processing circuitry 34, 35 may be of identical parameters and architecture, or be of different parameters or different architectures. The training of ML models may utilize the information about the location and / or orientation of the associated array. In this way, the ability of the system 400 to identify events may be modified to be suitable to the field of view of each array 30, 31 , 36.

[0086] Each ML model in the system 400 processes the electrical parameters received from the interface circuitry 32, 33, 37 and identifies an event (or, in some cases, a lack of an event or a persistent event state) in the field of view of its associated array 30, 31. Each processing circuitry 34, 35 then provides the event output by its associated ML model, in the form of a notification to an output device 39. While only one shared output device 39 is shown in Figure 5, there may be a plurality of output devices 39 used to display notifications from each respective processing circuitry 34, 35.

[0087] In this case, there are a smaller number of processing circuitries 34, 35 than interface circuitries 32, 33, 37. In some examples, the number of interface circuitries and processing circuitries may be equal. A plurality of arrays connected to the same processing circuitry may be used to track the movement of an event or object that is disposed in the fields of view of at least two of the arrays. In an example, an aircraft may travel on a path that intersects with the fields of view of at least two different arrays of photovoltaic devices. The ML model comprised in the processing circuitry that the at least two arrays are connected to may be trained to identify that it was indeed the same aircraft, hence tracking it across different fields of view. This could happen in two ways.

[0088] In a first case, the fields of view of at least two arrays may be physically separated as shown in Figure 4a. The processing circuitry would then identify two or more different events in two or more different fields of view. The processing circuitry may further identify that each event involved the same aircraft. The processing circuitry may further improve its confidence in the tracking by interpolating the path of the aircraft based on the speed and trajectory of the aircraft in the two or more fields of view.

[0089] In a second case, there may be one of more overlap regions between the two or more fields of view of the two or more arrays in the system. In this case, the processing circuitry would not need to interpolate the path of the aircraft in the physical space between the fields of view and would be able to track the aircraft with greater confidence. In some examples, the processing circuitry may be able to extrapolate the path of the aircraft beyond a field of view of an array.

[0090] Figure 6 is a schematic of a system 500 used to identify an event in a field of view of an array 40 of photovoltaic devices 20 and is independent of the systems described above. Features described in previous systems will not be described in detail in the description of the system 500. The description will only emphasise the differences between the system 500 and the system 400 of Figure 5. The system 500 comprises two arrays labelled 40 and 41 connected to interface circuitries 42 and 43 respectively. While only two arrays 40, 41 and associated interface circuitries 42, 43 are shown in Figure 6, there may be more than two of each. Each interface circuitry 42, 43 is connected to a single same central processing circuitry 44. While the arrays 40, 41 are collocated with the interface circuitry 42, 43, the processing circuitry 44 may be provided in a location remote from one, some or all of the arrays 40, 41 . The interface circuitries 42, 43 are electrically connected, via a wired connection, to the arrays 40, 41 . The processing circuitry 44 may be electrically connected to each of the interface circuitries 42, 43 via a wired or wireless connection. The processing circuitry 44 may be a hardware device or it may be hosted on the cloud. The processing circuitry 44 contains a ML model trained to identify one or more events in a field of view of the arrays 40, 41 provided in the system 500. The output of the ML model is provided by the processing circuitry 44 to the output device 46 in the form of a notification concerning the identified event or events to be output to a user.

[0091] The system 500 of Figure 6 may comprise arrays 40, 41 with fields of view that overlap or fields of view that are separate from each other or a combination of these two configurations. The single processing circuitry 44 will receive electrical parameters that are associated with the combination of the fields of view of the arrays 40, 41 connected to it. The ML model will hence be able to identify events in a plurality of fields of view. As a result, the processing circuitry 44 may be used to track events, temporally and / or spatially across the combined fields of view of all arrays 40, 41 connected to the processing circuitry 44 in the same way as described for the example of Figure 5. This will require providing the ML model with the electrical parameters as well as one or more of the time that the electrical parameters were generated by the photovoltaic devices 20, the location of the array 40, 41 and / or photovoltaic devices 20 when the electrical parameters were generated, and the orientation of the array 40, 41 and / or photovoltaic devices 20 when the electrical parameters were generated. Where the fields of view overlap and the aerial location of an event proceeds from one field of view to another, the system 500 may be able to track events continuously across the overlapping fields of view. Where the fields of view do not overlap and an event proceeds from one field of view to the other, such as the flight of a flock of birds, the system 500 may be able to track the event only within each relevant field of view. In such cases, the ML model may be further trained to interpolate the movement of the event in the space and time between the event being captured in the respective fields of view. By this mechanism, the system 500 can track events taking place in a combined field of view of a plurality of photovoltaic devices 20, spatially and / or temporally.

[0092] In an aspect of the disclosure, a processing apparatus is presented, the processing apparatus is configured to train a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices, the apparatus comprising training circuitry configured to train a machine learning model to perform event identification based on training data, wherein the training data comprises or is related to electrical parameters associated with events and wherein the electrical parameters are obtained from the array of photovoltaic devices. For example, the electrical parameters may be obtained from one or more (e.g. multiple, a plurality of or each) photovoltaic devices in one or more arrays.

[0093] In another aspect of the disclosure, a method of training a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices is presented, the method comprising: using optic apparatus to transform and select components of an input image, training the machine learning model to perform event identification based on training data, wherein the training data comprises or is related to electrical parameters associated with events and wherein the electrical parameters are obtained from the array of photovoltaic devices.

[0094] Experimental results

[0095] Figure 7 is a graph that shows characterization results of a fabricated perovskite solar cell, with a logarithmic voltage response against the incident light power. The measured voltage generated by the solar cell is plotted versus the input light power. The non-linear logarithmic response of the solar cell is evident in Figure 7 by comparing it to the logarithmic function. Figure 8 shows a schematic of an ONN proof-of-concept setup 600 built using a six-pixel solar cell array 56 formed of Perovskite solar cells and used to evaluate the ONN as it performs a classification task. The ONN was evaluated by performing MINST classification. The MNIST database (Modified National Institute of Standards and Technology database) is a large database of handwritten digits that is commonly used for training various image processing systems. The setup 600 was able to achieve a classification accuracy of 53%, and consistently outperformed a six-pixel charge-coupled device (CCD) equivalent.

[0096] In Figure 8, light from an input image 50 is encoded by a spatial light modulator (SLM) 52. The modulated light is then passed through a hidden layer 54 consisting of a scattering medium. The light is then captured by six separate solar cell pixels in the form of a two-dimensional planar solar cell array 56 and provided to a machine learning model 58, which uses trained weights to produce the classification results.

[0097] The SLM 52 in the setup 600 is a 512x512 liquid crystal SLM which was used to encode the light from the input image 50 into a collimated laser beam by using phase modulation. Undesired Fourier components were rejected using an optical iris (not shown) to improve the contrast of the encoded images by removing any background signal. The scattering medium used consisted of a layer of micron-sized alumina beads. The scattering medium function as randomized and fixed linear weights for the ONN. Even though the dimension of output feature space is highly restricted due to the use of a six-pixel solar cell array 56, significant improvement in system performance was achieved when adding the layer of scattering medium. When light having passed through the scattering medium is incident on the solar cell array 56, each solar cell pixel generates a non-linear voltage signal in response.

[0098] The physical propagation process for the setup 600 shown in Figure 8 can be formulated as where x(i) and h(i) are the i-th input image and its corresponding output state, p denotes the image encoding process, M is a selection mask, g is a randomized and fixed linear mapping carried out by the scattering medium, A is the intrinsically nonlinear acquisition operator of the solar cell, and f represents a computing reservoir formed by all optical components. For simplicity, the complex free-space transformation matrices in-between the optical elements are not shown in the presented formulation. Although the captured output h(i) can be trained for various downstream tasks, here we evaluate the ONN system for solving MNIST classification tasks by digitally training a single layer of linear weights, 0* that solves the regression problem where Y is the target label, Y = 0H is the system prediction, H = [h(1), h(2), ..., h(n)], and c is a regularization parameter. For low-dimensional H, the output weights 0* can be easily computed using: e* = HTH + CI)~1HTY

[0099] After training, this ONN system is able to achieve an average classification accuracy of 91 % on the reduced binary MNIST classification test set, and 53 % for a full ten-class test set.

[0100] For comparison, the same ONN setup has also been tested with a CCD camera instead of solar cells. CCD cameras have a linear response against the incident power until saturation, at which point the pixel readout will be capped at its maximum value. The physical propagation process can thus be reformulated as where the A' is a linear acquisition operator, s represents the saturation process, which is equivalent to taking min(f'(x(i)), 255), with 255 being the maximum pixel readout value of the CCD camera. Furthermore, to simulate the large pixels of the solar cells, the captured 1024 x 1024 CCD images were resized to 4 x 4, a size that gives the best performance, before selecting 6 pixels that carry the most information. This setup has achieved an accuracy of 88 % and 48 %, respectively, on binary and ten-class MNIST classification, consistently underperforming compared to its solar-cell equivalent, often by a sizable margin.

[0101] Figure 9 is a violin plot that compares the accuracy of the ONN setup using solar cells and those using CCD cameras, as described above. The number of classes corresponds to the number of different MINST digits the ONN is tasked to classify. Each violin illustrates the distribution of the accuracies evaluated over all of the unique combinations of a fixed number of classes, and the average accuracy of these combinations is annotated on the plot. Each violin is independently normalized by its area. This confirms that the solar cell setup consistently outperforms the equivalent CCD setup regardless of the number of classes. Whilst certain examples are described, these examples have been presented by way of example only, and are not intended to limit the scope of the disclosure. Indeed, the novel methods and systems described herein may be embodied in a variety of other forms. Furthermore, various omissions, substitutions and changes in the form of the methods and systems described herein may be made without departing from the scope of the disclosure. The accompanying claims and their equivalents are intended to cover such forms and modifications as would fall within the scope of the disclosure.

Claims

Claims1 . A computer implemented method for identifying an event in a field of view of an array of photovoltaic devices comprising: receiving, by processing circuitry, electrical parameters from one or more photovoltaic devices in the array, wherein the electrical parameters are dependent on electromagnetic wave propagation in the field of view of the array of photovoltaic devices; wherein the processing circuitry comprises a trained machine learning model configured to identify an event taking place in the field of view of the array based on the electrical parameters; and using the trained machine learning model to identify an event.

2. The method of claim 1 further comprising using interface circuitry to obtain the electrical parameters from the array of photovoltaic devices and to provide the electrical parameters to the processing circuitry.

3. The method of either preceding claim wherein the electrical parameters comprise one or more of measured voltage, current and output power obtained from the photovoltaic device.

4. The method of any preceding claim wherein the array comprises a plurality of photovoltaic devices arranged in one or more directions.

5. The method of any preceding claim wherein a field of view of the array comprises one or more solid angles through which the array can detect electromagnetic wave propagation.

6. The method of any preceding claim wherein the event identified in a field of view of the array is one or more of a meteorological event, movement of birds, an aerial surveillance event, lidar, optical communication and optical satellite communication.

7. The method of any preceding claim wherein the processing circuitry stores at least an indication of the event in a memory device and / or provides an identification of the event to the user in the form of a notification on an output device.

8. An apparatus for identifying an event in a field of view of an array of photovoltaic devices comprising processing circuitry configured to receive electrical parameters from one or more photovoltaic devices in the array, wherein the electrical parameters are dependent on electromagnetic wave propagation in the field of view of the array of photovoltaic devices; the processing circuitry comprising a trained machine learning model configured to identify an event taking place in the field of view of the array based on the electrical parameters.

9. The apparatus of claim 8 further comprising interface circuitry configured to obtain the electrical parameters from the array of photovoltaic devices and to provide the electrical parameters to the processing circuitry.

10. The apparatus of claim 9 wherein the interface circuitry is further configured to measure the electrical parameters from the array of photovoltaic devices.

11. A system for identifying an event in a field of view of an array of photovoltaic devices comprising: one or more arrays of photovoltaic devices; and processing circuitry configured to receive electrical parameters from one or more photovoltaic devices in the one or more arrays, wherein the electrical parameters are dependent on electromagnetic wave propagation in the fields of view of the arrays of photovoltaic devices; wherein the processing circuitry comprises a trained machine learning model configured to identify an event taking place in the fields of view of the one of more arrays based on the electrical parameters.

12. The system of claim 11 further comprising interface circuitry configured to obtain and / or measure the electrical parameters from one or more photovoltaic devices in the one or more arrays of photovoltaic devices and to provide the electrical parameters to the processing circuitry.

13. The system of claim 12 wherein the interface circuitry is further configured to determine one or more of a time associated with the received electrical parameters, the location of the one or more arrays of photovoltaic devices associated with the received electrical parameters and the orientation of the one or more arrays of photovoltaic devices associated with the received electrical parameters.

14. The system of any of claims 12 and 13 comprising a plurality of photovoltaic device arrays, wherein a plurality of the processing circuitries and a plurality of the interface circuitries are provided and each one of the plurality of interface circuitries is configured to provide the electrical parameters to a respective one of the plurality of the processing circuitries.

15. The system of any of claims 12 and 13 comprising a plurality of photovoltaic device arrays, wherein one or more of the processing circuitries and a plurality of the interface circuitries are provided and the plurality of interface circuitries are configured to provide the electrical parameters to the one or more processing circuitries.

16. The system of any of claims 12 and 13 comprising a plurality of photovoltaic device arrays, wherein a plurality of the interface circuitries are provided and the plurality of interface circuitries are configured to provide the electrical parameters to a number ofthe processing circuitries wherein the number of the processing circuitries is smaller than a number of the interface circuitries.

17. The system of any of claims 11-16 wherein the processing circuitry is configured to track events taking place in a combined field of view of a plurality of photovoltaic devices in the one or more arrays, spatially and / or temporally.

18. The system of any of claims 11-17 wherein the photovoltaic devices comprise one or more solar cells such as Perovskite solar cells, single junction solar cells and organic solar cells.

19. The system of any of claims 11-18 wherein at least one of the interface circuitries and the processing circuitries is powered by at least the photovoltaic devices.

20. The system of any of claims 11-19 wherein the system further comprises an output device configured to convey a notification of an identified event, received from the processing circuitry.

21. A processing apparatus configured to train a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices, the processing apparatus comprising training circuitry configured to train the machine learning model to perform event identification based on training data, wherein the training data comprises or is related to electrical parameters associated with events and wherein the electrical parameters are obtained from one or more photovoltaic devices in the array.

22. A method of training a machine learning model to identify an event taking place in a field of view of an array of photovoltaic devices, the method comprising: training the machine learning model to perform event identification based on training data, wherein the training data comprises or is related to electrical parameters associated with events and wherein the electrical parameters are obtained from one or more photovoltaic devices in the array.

Images